File size: 536 Bytes
0a138aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
default_stage:
  default_modifiers:
    SmoothQuantModifier:
      smoothing_strength: 0.8
      mappings:
      - !!python/tuple
        - ['re:.*q_proj', 're:.*k_proj', 're:.*v_proj']
        - re:.*input_layernorm
      - !!python/tuple
        - ['re:.*gate_proj', 're:.*up_proj']
        - re:.*post_attention_layernorm
      ignore: []
    GPTQModifier:
      targets: [Linear]
      ignore: [lm_head]
      scheme: W8A8
      sequential_update: true
      block_size: 128
      dampening_frac: 0.01
      offload_hessians: false